Van Huy

@jinfagang Sorry I don't have wechat. How about via email?

Actually I did. Generated images are pushed into a pool and used to update discriminators later. https://github.com/vanhuyz/CycleGAN-TensorFlow/blob/master/train.py#L97-L98 https://github.com/vanhuyz/CycleGAN-TensorFlow/blob/master/utils.py#L33

Yes, I will add identity mapping loss soon. But before that, I have to fix #18. Anyway, pull requests are always welcome! 😉

Actually, this is a trick which was described in the original paper (in `4. Implementation > Training details` section). > Second, to reduce model oscillation [15], we follow Shrivastava et...

Yes, just replace the pictures in data folder. Check `$ python3 build_data.py --help` to make sure how to specify the data folder :slightly_smiling_face:

Nice! Let me dig around it. About image pool, the current images are feeding 50% of the time as in: https://github.com/vanhuyz/CycleGAN-TensorFlow/blob/master/utils.py#L49-L57

I think this is a TensorBoard's bug...

It looks like the model has already collapsed... 🤔

I still can't figure out how to fix it, but adding *identity mapping loss* as the original paper suggested would somehow make a difference.

Depending on data, but I usually trained about 100k~200k steps. I think loss values in GAN do not mean too much. In my cases, value of D is very close...